Could Chinese researchers distill OpenAI models more sample-efficiently than OpenAI?

One poster hypothesizes that Chinese researchers may have covertly developed methods for distilling OpenAI models more sample-efficiently than OpenAI itself. A reply calls that impossible, arguing that OpenAI and Anthropic have more information through logits, the raw scores behind model outputs. The first poster says they mean end-to-end efficiency rather than gradient per token, and speculates that better ways to mix in trajectories from a stronger teacher model could help.